avifenesh/growth-brief

Monthly Growth Brief

Synthesizes one month of Cairn's operational data into a structured brief: top failure modes, what was learned, what improved, and 3 concrete proposals for Avi to approve.

Run on the 1st of each month at 08:00 UTC via cron, or triggered manually.

Step 1: Collect Raw Data

Run all four queries in parallel using cairn.shell with sqlite3 against ${DATABASE_PATH:-$HOME/.cairn/data/cairn.db}.

1a. Top recurring error patterns (30 days):

SELECT pattern_hash, pattern, count, category, first_seen, last_seen, auto_fixed
FROM error_patterns
WHERE last_seen >= date('now', '-30 days')
ORDER BY count DESC
LIMIT 10;

1b. Action exemplar performance (30 days):

SELECT action_type,
       COUNT(*) as total,
       ROUND(AVG(score), 2) as avg_score,
       SUM(CASE WHEN score > 0.6 THEN 1 ELSE 0 END) as successes,
       SUM(CASE WHEN score < 0 THEN 1 ELSE 0 END) as failures
FROM action_exemplars
WHERE created_at >= date('now', '-30 days')
GROUP BY action_type
ORDER BY total DESC;

1c. Experiment windows — what was tuned and what worked (60 days):

SELECT ew.window_type, ew.composite_score,
       e.parameter_name, e.variant_value, e.status, e.success_count
FROM experiment_windows ew
LEFT JOIN experiments e ON e.id = ew.experiment_id
WHERE ew.started_at >= date('now', '-60 days')
  AND ew.ended_at IS NOT NULL
ORDER BY ew.composite_score DESC
LIMIT 20;

1d. Session journal — task completion and efficiency (30 days):

SELECT
  COUNT(*) as total_sessions,
  ROUND(AVG(round_count), 1) as avg_rounds,
  ROUND(AVG(duration_ms) / 60000.0, 1) as avg_duration_min,
  SUM(CASE WHEN json_array_length(errors) > 0 THEN 1 ELSE 0 END) as sessions_with_errors,
  mode
FROM session_journal
WHERE created_at >= date('now', '-30 days')
GROUP BY mode
ORDER BY total_sessions DESC;

Also query memories for relevant learnings:

cairn.searchMemory: "improvement error fix recurring failure pattern"
cairn.journalSearch: "fixed improved resolved learned"

Step 2: Synthesize the Brief

Analyze the data and write a structured Telegram-ready brief. Keep it under 3000 characters total.

---

[MONTHLY GROWTH BRIEF — {MONTH} {YEAR}]

TOP FAILURE MODES (last 30 days) List the top 3-5 patterns from error_patterns by count. For each:

• Pattern (normalized): {pattern}
• Occurrences: {count} | Category: {structural/transient}
• Auto-fixed: yes/no | Status: resolved/active

WHAT CAIRN LEARNED Summarize key learnings from session_journal and action_exemplars:

• Task completion: {total_sessions} sessions, {avg_rounds} avg rounds, {avg_duration_min} avg min
• Best-performing action type: {action_type} (score {avg_score})
• Worst-performing action type: {action_type} (score {avg_score})
• Notable from journal: {1-2 sentence synthesis from cairn.journalSearch}

EXPERIMENT RESULTS Summarize what parameters were tuned:

• Parameters tested: {list from experiment_windows}
• Best window score: {composite_score}
• What improved: {parameter_name} = {variant_value} (adopted/rejected)

WHAT IMPROVED THIS MONTH 2-3 bullets on concrete improvements observed (compare to previous period if data exists).

3 PROPOSALS FOR AVI Concrete, actionable proposals based on the data. Each needs explicit approval before Cairn acts.

1. [HIGH] {Proposal title} — {1-2 sentence rationale from the data}
2. [MEDIUM] {Proposal title} — {rationale}
3. [LOW/EXPLORATION] {Proposal title} — {rationale}

---

Step 3: Send via Telegram

cairn.publishToFeed: send the formatted brief via preferred channel (Telegram)
Priority: medium (not urgent, but important)

Output Format Rules

• Keep total brief under 3000 characters (Telegram limit per message)
• If data is sparse (< 10 sessions, < 5 error patterns): note data is still accumulating and show what IS available
• Proposals must be concrete and measurable — not "improve performance" but "reduce CODING_MAX_ROUNDS from 100 to 60 to cut session cost ~40%"
• Never fabricate data — if a table is empty, say so explicitly
• Flag any patterns with count > 20 as HIGH PRIORITY in the failure modes section